SkillFoxx

Deepdive skill

A Claude Code skill that turns "research this topic" into a multi-phase investigation: hypotheses, parallel search, a file per source and citation checks

  • 367 ★ on the repository
  • Last commit
  • MIT
Skill

Extended access

Why this level

  • Launches sub-agents and Python scripts, calls dozens of external sites and APIs and writes files into the research folder
  • A deep run spends many tokens, and keys for paid APIs are passed through environment variables
All reasons and checks

socialpranker/deepdiveAuthor:Socialpranker

Install

Manual install

git clone https://github.com/Socialpranker/deepdive.git ~/.claude/skills/deepdive

After cloning, the skill fires on phrases like "Investigate X" and "Изучи тему".

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • Launches sub-agents and Python scripts, calls dozens of external sites and APIs and writes files into the research folder
  • A deep run spends many tokens, and keys for paid APIs are passed through environment variables

This is third-party code. Review the repository files before installing.

What it does

Deepdive guides Claude Code through research in a strict order of 13 phases and leaves behind a folder you can return to in a month. The skill fires on phrases like "investigate X", "изучи тему" and "validate this hypothesis". It first reframes the question, forms hypotheses and records a decision spec: what you will do with the answer and under what condition you pick option A. Then it chooses one of six report genres (Q&A, explainer, decision, landscape, validation, custom) and writes plan.md. Before search a plan-review gate opens: you see the framing, hypotheses, genre and channels and can edit them. Search runs across 29 named channels, from academic databases to forums and news, with sub-agents working in parallel. Every source found is saved as its own file with verbatim quotes. The skill tracks provenance: ten articles retelling one press release count as one voice. Next come claim triangulation across sources, an evidence filter and a red team that attacks the conclusions in four roles. The output is a report with every claim tied to a quote, a one-page decision memo and a walkthrough of the decision forks one by one. The repository ships scripts for checking: check_citations.py verifies that links are alive, validate_phases.py confirms every mandatory phase left a file, and number provenance and arithmetic checks catch invented figures. The skill announces up front which phases run on which models and gives an approximate cost.

Who it is for. For analysts, founders and developers who need reproducible desk research with citations and source checks before a decision, rather than a one-off chat answer.

Good fit when

  • You must compare approaches before choosing, such as Postgres logical replication and CDC tooling, and record which option you picked and why
  • You need to test a hypothesis against external data and get a report where every claim points to a file with a quote
  • A strategy, talk or article is in preparation and you want a folder with plan, sources and theses to come back to
  • You need a landscape map: who is in the market and on what criteria to compare them
  • You want to check a finished report for dead links, swapped quotes and invented numbers with the repository's scripts

Not a fit when

  • You need a quick fact check: the agent should simply answer instead of starting a chain of phases
  • The question concerns the Anthropic SDK or Claude API: SKILL.md itself routes such requests to another skill
  • You want brainstorming without data or a competitor comparison against a fixed matrix: the authors point to other skills for that
  • The token budget is tight: a deep run involves up to five sub-agents and 25 to 35 sources, and the skill announces the cost up front

Example request

Investigate the trade-offs between Postgres logical replication and CDC tooling and help me pick an option for our project

Limitations

  • The main flow is built for Claude Code and its sub-agents. For Codex, Gemini and local models the method can be ported by hand by loading SKILL.md and the references into context, without parallel sub-agents.
  • The SKILL.md description and triggers are written in Russian and English, while the references are mostly English.
  • A deep run costs noticeable tokens: the authors estimate it at about 2 dollars instead of 8 thanks to routing models by phase. The figure depends on the topic.
  • Some API channels are free without a key, and keys for the rest are read only from environment variables; the skill never asks for them in chat.
  • The catalog of more than 47 APIs is validated weekly by automation, but stale access terms are possible, so each entry carries a verification date.

How to disable. Delete the ~/.claude/skills/deepdive folder and the skill stops firing. Research folders it already created with plans and sources stay as ordinary files in your projects.

FAQ

How do I install Deepdive in Claude Code?

Run git clone https://github.com/Socialpranker/deepdive.git ~/.claude/skills/deepdive. After that a phrase such as "Investigate X" or "Изучи тему" in a session is enough. For Claude Desktop, zip the repository into a .skill file and upload it under Settings, Skills.

How is Deepdive different from a plain "search the web" request?

Instead of one answer the skill goes through phases: it reframes the question, shows a plan for approval, saves every source as its own file with quotes and always runs a red team. The result is a folder with a report where every claim points to a specific quote.

How much does a deep Deepdive run cost?

The skill announces a cost estimate once at the start and names which phases run on which models. By the authors' estimate, routing models by phase brings a deep run to about 2 dollars instead of roughly 8. You can override it with the phrases "with all on opus" and "with cheap mode".

How does Deepdive check sources and citations?

Checks run in four layers: is the link alive, does the source entail the claim, are the qualifiers preserved, and are the framework names invented. Separate scripts check number provenance and arithmetic. The scripts live in the repository and run as check_citations.py and validate_phases.py.

Can I use Deepdive without Claude Code?

Yes, with losses. The 13-phase method is portable: load SKILL.md and the files under references into the context of Codex, Gemini or a local model. There will be no parallel sub-agents, so subtopics are researched in separate chats.

Discussion

No questions or reviews yet.

Skills

Deep Research

Deep Research Skill for Claude Code

A deep research skill for Claude Code: multi-phase search, claim verification and cited reports in Markdown, HTML and PDF

SkillExtended access1.2K

A Claude Code skill that turns the agent into a researcher: multi-step source collection, citation verification, and a wiki report that grows session by session

SkillExtended access4.7K

A skill for Claude Code and Codex: one query goes to 52 search channels and the results merge into one deduplicated JSON

SkillExtended access772

A set of skills for searching scientific databases, PubMed, arXiv, ChEMBL, ClinicalTrials and more, with plain-language queries.

PluginExtended access64

Google Scholar Skills

Google Scholar Skills for Claude Code

Six skills and a gs-researcher agent: Claude Code searches Google Scholar, tracks citations, fetches full-text links and exports to Zotero

SkillExtended access545

A skill and CLI for agents: search and reading across 23 channels, from web and GitHub to social networks, with an evidence rule.

SkillBroad access60

Comparison with similar tools

Cards that share the main search query, and how they differ.

ToolOfficialRepository starsMain difference
Deepdive (this card)no367
Deep Research Skillsno2,550a structured deep-research skill with human control at the stages
Research Toolkitno314a skill that takes research from a question to a report where every conclusion has a source
GPT Researcherno30,652an autonomous deep-research agent with no phases or source files

README in short

The repository page shows a sample dialogue and a comparison with an ordinary request, then three install methods and a table of 13 phases with a model for each. It goes on with details of search, source scoring, citation checks and the decision memo, catalogs of report blocks, search channels and APIs, plus the checking scripts and the weekly automated sync.

README badge

Original SKILL.md

---
name: deepdive
description: "Meta-research под вопрос или решение: веб-поиск, источники, Q&A отчет с цитатами по файлам для повторного использования. Использовать для деск-ресерча, валидации гипотезы, «как устроен X». Триггеры: «deepdive», «сделай ресерч», «исследуй», «копни глубоко», «ресерчни»."
---

# Deepdive — meta-research с дисциплиной

Многошаговое исследование под вопрос или решение. Источник = файл, отчет как Q&A, тезисы атомарны и пере-используемы.

## Когда применять

Прямая просьба о ресерче; сравнение N институций/продуктов/методологий/рынков; материал под стратегию, доклад, статью; проверка гипотезы внешними данными; «как устроен X», «карта области Y».

**НЕ применять:** быстрая фактоверка → отвечай напрямую · N конкурентов по фиксированной матрице → `competitive-teardown` · Anthropic SDK / Claude API → `claude-api` · брейншторм без данных → `brainstorming`/`grill-me` · ответ уже в проекте → сначала grep.

## Глубина — по теме, жесткого дефолта нет

| Режим | Источников | Суб-агентов | Когда |
|---|---|---|---|
| shallow | 5–7 | 0 | первичная навигация, тема знакома, low-stakes |
| medium | 12–18 | 2–3 | нетривиальная тема, среднее решение |
| deep | 25–35+ | 4–5 | high-stakes решение, стратегия |

Объяви режим в начале с обоснованием. После Genre+Plan объяви **model routing** одной строкой (фазы по моделям + estimated cost + как перебить: «все на opus» / «cheap mode»). Детали — `model_routing.md`.

## Перед стартом — discover existing

До reframing (опционально — нет файлов, иди дальше): определи целевую папку → есть — перечисли содержимое, похожий slug ⇒ спроси «это update?» → прочитай `CLAUDE.md`/`CLAUDE.local.md` и `memory/MEMORY.md`, учти в reframing. Цель: не дублировать сделанное.

**Плюс кросс-прогонная вики** — `python scripts/wiki_query.py --topic "<вопрос>"`: прошлые утверждения, уже оцененные источники (credibility не пересчитывать), открытые противоречия между прогонами. Непогашенное противоречие идет в `plan.md` исследовательским вопросом. См. `wiki.md`.

**Куда сохранять** (не хардкодь): (1) research-папка из CLAUDE.md или существующая `research/` · `06_Деск-ресерч/` · `docs/research/` · `notes/research/`; (2) иначе по типу проекта — манифест (`pyproject.toml`/`package.json`/`Cargo.toml`/`go.mod`) → `research/`, только документы → `06_Деск-ресерч/`; (3) не git-репо или пусто → `~/deepdive/<slug>/`. Путь покажи ОДИН раз, дальше пиши молча.

**Slug:** латиница, цифры, дефисы («Postgres logical replication vs CDC» → `postgres-replication-vs-cdc`). Неочевиден — покажи в начале фазы 2.

## Workflow — <!--gen:count:phases-->13<!--/gen--> фаз (1–8, включая 3.5, 3.7, 5.5, 5.7, 6.5)

Детали фаз — `workflow.md`, модель на фазу — `model_routing.md`. Здесь — что фаза обязана оставить после себя.

1. **Reframing** [`opus`/high] — переписать вопрос; **Decision Spec** (решение глагол+объект+срок / потребитель→его следующий шаг / ≥1 if-then вилка «покажет X → делаю A»; ни одной вилки ⇒ честный даунгрейд в shallow); 2–4 опровергаемые гипотезы; medium/deep — персоны охвата (STORM) и router по типу вопроса. См. `question_reframing.md`.
2. **Genre & blocks** [`sonnet`/medium] — жанр (qa/explainer/decision/landscape/validation/custom) + набор блоков, подтвердить одной строкой. См. `genres.md`, `blocks/INDEX.md`.
3. **Plan** [`opus`/medium] — `plan.md` по шаблону §0–16 из `workflow.md`: HEADER → SCOPE → STRUCTURE → EXECUTION → TRACKING. Несущее: acceptance criteria, гипотезы, risk register, subtopic↔blocks mapping (least-to-most для многошаговых вопросов), sourcing strategy §12, opposition queries, stop-criteria.
3.5. **Capability Discovery** [`sonnet`/low] (deep — обязательна) — audit env vars, подтемы → доступные API, fallback на awesome-lists. См. `capability_discovery.md`.
3.7. **Plan-review gate** [`sonnet`/low] (shallow — skip) — единственная human-in-the-loop точка ПЕРЕД дорогой Фазой 4: показать сжатый план (вопрос, решение, жанр, гипотезы, каналы, стоп-критерий, routing). **deep — ЖДАТЬ явного «Ок»; medium — soft.** Плюс **скаут-пасс** (deep — рекомендуется): 3–4 `Explore` на `haiku` ищут непокрытые подвопросы, а не источники; выход — правки `plan.md`, ноль записей в `sources/`. См. `plan_gate.md`.
4. **Поиск** [main `sonnet`/medium; sub-agents: `haiku` web/api, `sonnet` academic/long-source] — (4.0) Source Dispatch по матрице → `plan.md` §12; количественный подвопрос ⇒ primary-канал registry/API. (4.1) medium/deep — `general-purpose` суб-агенты параллельно, каждому свой диапазон id (`s01-s09`, `s10-s19`…) и своя ось поиска, не только подтема; shallow — главный поток. (4.2) Fetch, дедуп с замером `overlap_rate`; WebFetch вернул тонко (<500 знаков, нет цитаты, блок/paywall) ⇒ сразу `fetch_source.py`, не по желанию. (4.3) Агент сам пишет `sources/NN_slug.md`, в главный поток — только index-строки. После раунда 1 — snowball. Loop: goal-check → bounded deviation → circuit breaker (2 раунда без нового ⇒ стоп, остаток в Open Questions). **Окно раунда пересобирается, не накапливается** (medium/deep): `state.md` ПЕРЕЗАПИСЫВАЕТСЯ перед каждым раундом (`## Known` статусами со ссылками · `## Gaps` · `## Next`, ≤6 КБ), планирование по нему, а не по транскрипту. См. `source_dispatch.md`, `subagents_v2.md`.

Report a problem